JPH1195795A - Voice quality evaluating method and recording medium - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve the reliability by giving a high score to a correctly uttered voice and giving a low score to an erroneously uttered voice. SOLUTION: A correct solution absolute logarithmic likelihood distribution PGP1 (X), which is defined as the distribution of the likelihood value between phoneme segments and the probability model of correct solution phonemes, and an incorrect solution absolute logarithmic likelihood distribution PGP1 P2 (X) are beforehand obtained by employing learning voices and the distributions are partitioned into segments against evaluation voices and correct solution phoneme groups are assigned (S1). Then, a likelihood SGP1 =x, which is the model of a correct solution phoneme P1 of each segment, is obtained (S2). Then, beforehand-obtained PGP1 (x) and each PGP1 P2 (x) are obtained by substituting (x) and the sum value of each difference is defined as an accumulated absolute score SCP1 (X)(S3).

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method and a recording medium for evaluating the quality of uttered speech using a computer.

[0002]

2. Description of the Related Art As a conventional speech quality evaluation method, there is a method based on a spectrum distance scale (for example, Reference 1, Kenzo Ito, Nobuhiko Kitawaki, Kazuhiko Kakehi, "Study of objective evaluation scale of digital waveform coding method for speech". ”, IEICE,
Vol. J66-A, No. 3, pp. 273 (198
3)). For example, when evaluating speech through a certain communication system or coding method, the quality of speech can be evaluated by comparing the spectrum of the original speech at the input end with the spectrum of the speech at the output end. In this method, the original voice before the evaluation voice is distorted is required as the reference (reference) voice. Therefore, it cannot be applied to the case where the voice of a foreign language utterance trainer is evaluated. In language utterance training, the reference voice is the voice of the language trainer who has acquired the same utterance method as that of the native speaker, and cannot be obtained during training. Although it is conceivable to use the voice of an arbitrary native language speaker as a reference, since the voice has a speaker-specific spectral characteristic, it may be the difference between the spectral characteristics due to the immaturity of the utterance, There is a problem that correct evaluation cannot be performed because it is not possible to distinguish whether the difference is due to a difference in spectral characteristics due to the difference between. To solve this problem, a hidden Markov model (Hidden Ma
There is a voice quality evaluation method using a probabilistic model such as rkov Model (rkov Model: HMM). Since the HMM is obtained by learning using the voices of a large number of speakers, various speaker characteristics can be included in the model. If a large amount of learning data is prepared, the influence of the difference between speakers can be reduced, and only the difference due to the difference in utterance method can be reflected in the evaluation. For example, in order to apply the method using the HMM to speech evaluation during foreign language utterance training, speech of a large number of native speakers is collected, and an HMM is created based on the collected speech. Using this as a reference, evaluation is performed by comparing with the voice of the trainee.

[0003] Details of the HMM are described in, for example, Reference 2 (Seiichi Nakagawa:
Probabilistic model-based speech recognition, IEICE). FIG. 7A shows an example of a three-state continuous mixture distribution HMM. Such a model is created for each voice unit (phonemes, syllables, words, etc.). To each of the states S1 to S3, a statistical distribution D1 to D3 of audio feature parameters is given. For example, if this is a phoneme model, the first
The state represents the statistical distribution of the feature amount near the beginning of the phoneme, the second state near the center, and the third state near the end.

The feature distributions D1 to D3 in each state are often expressed using a plurality of continuous probability distributions (hereinafter, referred to as mixed continuous distributions) in order to express a complicated distribution shape. Although various distributions can be considered as the continuous probability distribution, a normal distribution is often used. Further, each normal distribution is often represented by a multidimensional uncorrelated normal distribution having the same number of dimensions as the feature amount. FIG. 7B shows an example of the continuous mixture distribution. In this figure, the number of normal distributions is 3 of N1 to N3.
One. Input feature vector ot = (ot1, t2,..., OtP) at time t in state i of the mixed continuous distribution HMM.
The output probability bi (ot) for (P is the total number of dimensions) is calculated as follows: bi (ot) = Σ _{m = 1} ^M wim · φi, m (ot) (1) Here, wim represents a weight coefficient for the m-th multidimensional normal distribution in state i. The probability density for the multidimensional normal distribution m is

[0005]

(Equation 1)

[0006] It is calculated as follows. Here, μim represents a mean vector for the m-th multidimensional normal distribution in state i, and Σ _im represents a covariance matrix. T represents the transpose of the matrix. Assuming that the covariance matrix of each normal distribution has only diagonal components, the logarithmic value of φim (ot) is

[0007]

(Equation 2)

[0008] Here, μimp is the p-th component of the average vector of the m-th multidimensional normal distribution of the state i, and σimp ² is the p-th component of the covariance matrix of the m-th multidimensional normal distribution of the state i. Represents a diagonal component (variance value). As a score in the conventional speech evaluation method using the HMM, the log likelihood value of Expression (3) was used. HM
Since M includes the features of the voices of many speakers as a probability distribution, it is possible to perform more appropriate evaluation than the method using the voice of one speaker as a reference. However, in the conventional method using only the log likelihood value calculated by Expression (3), a low log likelihood value may be shown for a correctly uttered voice, which is inappropriate as an evaluation value. was there.

[0009]

SUMMARY OF THE INVENTION An object of the present invention is to provide a voice quality evaluation method using a stochastic model, in which a high score is given to a voice in which a vowel consonant or the like is uttered correctly, and a voice having an incorrect utterance is given. It is an object of the present invention to provide a highly reliable voice evaluation method that gives a low score to the user.

[0010]

According to the present invention, an absolute score and a relative score for an evaluation voice are calculated using a probability model relating to a voice feature amount of a basic voice unit such as a vowel or a consonant. Quality evaluation is performed using either or both of the above. First, a procedure for obtaining an absolute score using a phoneme as a basic speech unit will be described. At the time of learning, a large number of training speech samples are decomposed into phonemes such as vowels and consonants to obtain phoneme segments, and then the likelihood of each speech segment and the probability model of the correct phoneme (correct linguistic category) to which it belongs. A degree value, that is, a likelihood that is a correct category, for example, a log likelihood value (referred to as a correct absolute log likelihood) is calculated. Next, a distribution for storing information on how the correct absolute log likelihood value is distributed, that is, a distribution of a correct category and log likelihood, and (correct absolute log likelihood distribution) is obtained. deep.
Further, the likelihood value (for example, log likelihood value, non-correct log likelihood value) of each speech segment and a probability model of an erroneous phoneme (erroneous linguistic category) different from the utterance is calculated. As for the non-correct absolute log likelihood value, a distribution for storing information on how the distribution is distributed, that is, a distribution of the likelihood as a wrong category (non-correct absolute log likelihood distribution) is obtained. . At the time of evaluation, a speech to be evaluated and a phoneme sequence to be uttered are input and decomposed into phoneme segments. For each phoneme segment, the correct absolute log likelihood is calculated, and the likelihood value indicating the likelihood obtained from the correct absolute log likelihood and the correct absolute log likelihood distribution, that is, the likelihood that the calculated likelihood is the correct answer. (Likelihood 1) and the likelihood obtained from the correct absolute log likelihood and the non-correct absolute log likelihood distribution, that is, a value indicating the likelihood that the calculated likelihood is a non-correct answer (likelihood 2). Is calculated. An absolute score is obtained from the obtained two likelihoods (likelihood 1 and likelihood 2).

Next, a procedure for obtaining a relative score will be described. During training, a large number of training speech samples are used to determine the likelihood value (eg, the absolute log likelihood of the correct answer) between each speech sample and the probability model of the correct language to which it belongs, and the wrong phoneme (the wrong linguistic category ) And the likelihood value (eg, non-correct absolute log likelihood) with the probability model (correct relative log likelihood)
Is calculated. Next, a loose distribution that stores information about the extent to which the correct logarithmic likelihood values are distributed, that is, the distribution of the difference between the likelihood of a correct category and the likelihood of a wrong category, Relative log likelihood distribution)
Ask for. Further, the likelihood value (non-correct relative log likelihood) of the difference between the log likelihood value (non-correct absolute log likelihood) of each speech segment and the erroneous phoneme probability model is calculated. Regarding the non-correct relative log likelihood value, the distribution for storing information on how wide it is distributed and the distribution of the difference between the likelihoods of each incorrect category (non-correct relative log likelihood distribution) Ask for it. At the time of evaluation, a speech to be evaluated and a phoneme sequence to be uttered are input and decomposed into phoneme segments. The correct relative log likelihood is calculated for each phoneme segment, and the correct relative log likelihood and the likelihood obtained from the previously stored correct relative log likelihood distribution, that is, the correct category and the non- A value indicating the likelihood that is different from the correct category (likelihood 3), the likelihood obtained from the correct relative log likelihood and the previously stored non-correct relative log likelihood distribution, that is, the calculated likelihood, A value indicating the likelihood of a difference between the target non-correct answer categories (likelihood 4)
Is calculated. A relative score is obtained from the obtained two likelihoods (likelihood 3 and likelihood 4).

The most main feature of the final evaluation score is to employ a score calculated from either the absolute score or the relative score or both. Conventional methods score the correct absolute log likelihood value of the evaluated speech and the probability model of the correct language category to which it belongs. Therefore, the method of the present invention considers the log likelihood value for the probability model of the wrong language category, the point of considering how the log likelihood value is distributed, only the absolute likelihood value. However, it differs from the conventional method in that a relative likelihood value is considered.

In the above probability model, the use of the HMM makes it easy to estimate model parameters, and the estimation accuracy is high. In addition, the above-described correct absolute log likelihood distribution, non-correct absolute log likelihood distribution, correct relative log likelihood distribution, non-correct relative log likelihood distribution, a normal distribution or a sigmoid function that is mathematically easy to handle is used. Easy to find distribution parameters.

[0014]

The absolute log likelihood value of the correct phoneme category and the probability model used in the conventional evaluation method indicates the similarity to the learning speech. of course,
This absolute index is important in evaluating speech. However, in a phoneme category such as a vowel consonant, there are a plurality of similar categories (for example, phonemes / b /, / d /, / g /), so that the evaluated voice is a sound close to other categories. It is important to know whether or not to evaluate (using non-correct absolute log-likelihood distribution). In the embodiment of the present invention, an absolute score considering both the correct absolute log likelihood distribution and the non-correct absolute log likelihood distribution is used. For example, phoneme /
If there is a voice sample of b /, the log likelihood value for the probability model of phoneme / b / is high,
If the log likelihood value for the stochastic model of d / is high, it should be difficult to say that this sample is the correct phoneme / b /. The absolute score of the present invention can obtain an evaluation value in consideration of such an event.

In some cases, the absolute log likelihood value is low even though the utterance is correct. This is because the features (for example, cepstrum) of speech used in speech recognition are
This is because it is appropriate for discriminating speech (determined only by the magnitude relationship between the phonemes), but may not be appropriate for absolute evaluation. Although the absolute log likelihood value with the probability model of the correct phoneme category is low, the absolute log likelihood value with respect to the probability model of another category is often lower.
Even if the absolute log likelihood value is low, when it can be said that it is overwhelmingly close to the correct phoneme category, it is considered that it can be regarded as a correct utterance. In the embodiment of the present invention, the difference (relative score) between the absolute log likelihood value with the probability model of the correct phoneme category and the absolute log likelihood value with the probability model of the wrong phoneme category.
By taking into account the above, it is possible to obtain an evaluation value in consideration of the above events.

The overall score is determined in consideration of one or both of the absolute score and the relative score. When both are considered, the absolute sound quality evaluation described above and the relative sound quality evaluation with respect to other categories can be performed, and the highly reliable voice quality evaluation, which is the object of the present invention, can be performed. Become like

[0017]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment in which the present invention is used for foreign language utterance training will be described. Here, the probability model is H
MM was used. First, the learning step (system preparation stage) will be described. The HMM for an unspecified speaker for each phoneme is learned based on a speech database uttered by a native language speaker. For example, when the target language is Japanese, all phoneme HMMs appearing in Japanese are learned using voices uttered by Japanese. Considering the acoustic effects of the phonemes before and after,
Create a phoneme environment dependent HMM. FIG. 1 is a flowchart for obtaining a correct absolute log likelihood distribution and a non-correct absolute log likelihood distribution. A learning speech sample and a phoneme sequence based on the utterance are given. Concatenate phoneme HMMs according to phoneme sequence,
Matching with a voice sample, a segment voice corresponding to each phoneme is obtained (S1). A log likelihood value between the HMM of the correct phoneme and the phoneme segment is obtained (S2). The log likelihood value is normalized according to the length L. The normalization method is, for example, as follows: Sno = Ss / L (4) Here, Ss is a log likelihood value obtained from a phoneme segment, and Sno is a normalized log likelihood value. Here, the normalized likelihood is referred to as a correct absolute log likelihood.

The correct absolute log likelihood S ^G _P1 is an absolute log likelihood value normalized by a segment length when a correct phoneme model (phon1) is given. Further, the log likelihood value of the wrong phoneme with respect to the HMM is calculated for the phoneme segment. Similarly, the log likelihood value is normalized by the segment length.
This is referred to herein as non-correct absolute log likelihood.

The non-correct absolute log likelihood S ^N _P1P2 is a phoneme (ph
on1), a competing phoneme model (pho
This is the absolute log likelihood value normalized by the segment length when n2) is applied. Phoneme with learning data (ph
For on1), after calculating the correct absolute log likelihood and the non-correct absolute log likelihood, the following cumulative distribution function is drawn (S4, S5), for example, a curve as shown in FIG. 2A is obtained.

F (x) = Prob ｛S ^G _P1 <x｝ (5) g (x) = Prob ｛S ^N _P1P2 <x｝ (6) A sigmoid function that can easily handle these cumulative frequency distributions mathematically. Approximate (S6, S7). f (x) = 1 / {1 + exp (−α1 (x−β1))} (7) g (x) = 1 / {1 + exp (−α2 (x−β2))} (8) where α and β Is 0.1 to 0.9 of the sigmoid curve
Is determined to be the best fit between the ranges of The probability density distributions of these functions are given mathematically as follows:

[0021]

(Equation 3)

[0022]

(Equation 4)

FIG. 2B shows an example of the distribution given by the above probability density function. The horizontal axis is the likelihood. These are the correct absolute log likelihood distribution and the non-correct absolute log likelihood distribution. FIG. 3 shows the cumulative distribution function when the correct absolute log likelihood is obtained for the phoneme (phon1), and the non-correct absolute log likelihood is obtained for every counterphoneme (phon2: 25 phonemes in this case). Accordingly, there is one function in equation (9) for each phoneme, and there are as many functions in equation (10) for each phoneme as the number of opposing phonemes.

FIG. 4 is a flowchart for obtaining a correct relative log likelihood distribution and a non-correct relative log likelihood distribution. In the same manner as in the case shown in FIG. 1, the training speech sample and each phoneme segment are decomposed, and a corresponding correct phoneme of a correct phoneme sequence is assigned to each phoneme segment (S1). For each of these segments, the correct absolute log likelihood S ^G _P1 is obtained (S2), and the non-correct absolute log likelihood S ^N _P1P2 is obtained (S3). The difference between the correct absolute log likelihood and the non-correct absolute log likelihood is defined as the correct relative log likelihood and is calculated as follows (S4).

Δ ^G _P1P2 = S ^G _P1 −S ^N _P1P2 (11) The non-correct relative log likelihood is calculated as follows (S5). _{^{Δ N P1P2 = (S N P1}} ) '- S N P1P2 (12) (S N P1)' = (1 / (n-1)) is Σ ^N _P1P2 (13) where sigma does not match the phonemes (phon1) Each phoneme (p
is the sum of S ^N _P1P2 N for Hon2) is the total number of phoneme model. For the correct relative log likelihood and the non-correct relative log likelihood, the respective cumulative distribution functions are drawn in the same manner as the absolute score (S6, S7), and after applying the sigmoid function (S8, S9), the probability density function ( q
^G _P1P2 (x), q ^N _P1P2 (x)) are obtained (S10, S1
1). These are the correct relative log likelihood distribution and the non-correct relative log likelihood distribution. The above is the learning step.

Next, an evaluation step of giving an evaluation value when a speech segment to be evaluated is given will be described. FIG. 5 is a flowchart for obtaining an overall score. The phoneme sequence of the speech to be evaluated and the target speech to be uttered are given. The phonemes HMM are connected according to the phoneme sequence, and the input speech is segmented into phonemes (S1). Correct phoneme (p
correct absolute logarithm likelihood S ^G _P1 = x for Hon1) (S2), the formula (9), calculates the likelihood value in accordance with equation (10). The overlap of the two distributions in FIG. 2B is an error area where an erroneous determination is made. The following can be considered as a score function reflecting this overlap.

[0027] ^{_{D P1P2 (x) = P G}} P1 (x) -P N P1P2 (x) (14) wherein, x is the absolute log-likelihood correct. When D _P1P2 (x) is positive, it is regarded as a correct utterance, and when it is negative, it is regarded as an erroneous utterance. The final absolute score SC _P1 (x) is calculated as follows (S3). SC _P1 (x) = _{{D P1P2} (x) (15)} is the sum of D _P1P2 (x) for all phon2s that do not match phon1.

Next, for each phoneme segment, the non-correct relative log likelihood S _{P1 P2} for the non-correct phoneme is determined (S
4). Based on the correct absolute log likelihood and the non-correct absolute log likelihood, the correct relative log likelihood is calculated according to equation (11) (S5). As in the case of the absolute score, the final relative score ΔSC _P1 (x) is calculated as follows from the correct relative log likelihood distribution and the non-correct relative log likelihood distribution (S6).

Δ _P1P2 (x) = q ^G _P1P2 (x) −q ^N _P1P2 (x) (16) ΔSC _P1 (x) = ΣΔ _P1P2 (x) (17) Here, Σ represents phon2 other than phon1. Δ
_P1P2 (x), where x is the correct relative log likelihood. The total score SCto is obtained from the absolute score SC and the relative score ΔSC. For example, consider a linear combination of both (S7).

SCto = λ · SC + (1−λ) · ΔSC (18) where λ is a coupling coefficient. When λ is 0, only the absolute score is reflected, and when λ is 1, only the relative score is reflected in the overall score. The combination of the two scores is not limited to a linear function, but may be a non-linear function. In the above examples, evaluation was made for each phoneme. When this is evaluated on a word or sentence basis, a weighted average of the overall score for each phoneme or the like can be considered.

FIG. 6A is an example of the result of configuring an HMM using mel-cepstral as a speech feature and applying the evaluation method according to the present invention. This figure shows multiple phonemes / b
This is the result when the voice sample of / is evaluated as phoneme / b /, and each point is the result for each sample. This experiment assumes that the correct utterance was made. The horizontal axis is the absolute score, and the vertical axis is the relative score. FIG. 6B shows a result when voice samples other than a plurality of phonemes / b / are evaluated as phonemes / b /. This experiment assumes that a wrong and inappropriate utterance was made. From these results, in the case of a correct utterance (sample), in most cases, both the absolute score and the relative score are large, but there are many cases in which only the relative score is large. In the case of an incorrect utterance (sample), both the absolute score and the relative score are often small, but the absolute score is not always small. Here, the effect of introducing a relative score was observed.

If a linear function is used to obtain the overall score, as shown in FIGS. 6A and 6B, points on a straight line have the same score in a two-dimensional space composed of absolute scores and relative scores. Is calculated as If a non-linear function is used, points on the curve can be given so as to have the same score, so that a more appropriate evaluation value can be given to the sample distribution.

In the above description, the probability model using a phoneme as a unit has been used, but other basic speech units such as syllables and words may be used. Also, using the absolute log likelihood, the correct likelihood for the probability model of the category of the correct speech unit, the correct likelihood distribution of how the correct likelihood shows, and the non- The likelihood distribution and the likelihood distribution of the correct likelihood and how the non-correct likelihood shows may be obtained.

[0034]

As described above, when an acoustic model such as an HMM for voice recognition is used, an absolute score may be bad even with correct utterance. However, even in that case, the relative score is often high. Therefore, if both the absolute score and the relative score are used, more accurate utterance quality evaluation can be performed.

The present invention can be used, for example, for vocal training in a foreign language.

[Brief description of the drawings]

FIG. 1 is a flowchart showing a procedure for generating a correct absolute log likelihood distribution and a non-correct absolute log likelihood distribution from learning speech samples.

FIG. 2A is a diagram showing an example in which each cumulative degree distribution of a correct absolute log likelihood and a non-correct absolute log likelihood is approximated by a sigmoid function;
B is a diagram showing an example of these probability density functions.

FIG. 3 is a diagram illustrating examples of systematic distribution functions of a correct absolute log likelihood and each non-correct absolute log likelihood.

FIG. 4 is a flowchart showing a procedure for generating a correct relative log likelihood distribution and a non-correct relative log likelihood distribution from learning speech samples.

FIG. 5 is a flowchart showing an example of a procedure of quality evaluation according to the method of the present invention.

FIG. 6 is a diagram showing an embodiment of the present invention.

7A is a diagram illustrating an example of a three-state HMM, and FIG. 7B is a diagram illustrating an example of expressing a probability distribution of the HMM;

Claims (13)

[Claims] 1. Using a plurality of linguistic categories of basic units such as phonemes, syllables, and words represented by a probability model related to acoustic features, and a likelihood of being a correct category calculated from the probability model in advance by learning speech. Inputting a voice sample to be evaluated and a basic unit sequence to be uttered corresponding to the voice, and calculating the voice unit as a basic unit. Decomposing into a basic unit; applying a probability model of a correct category to the decomposed basic unit to obtain a likelihood; distribution of the obtained likelihood and the likelihood of the correct category; From the distribution of likelihoods that fall into categories, a value that indicates the likelihood that the likelihood is correct and a value that indicates the likelihood that each likelihood is incorrect are obtained. -Up and voice quality evaluation method and a step of calculating an absolute score for evaluating quality from the value indicating these certainty. 2. The method according to claim 1, wherein the distribution of the likelihood as a correct category is a correct absolute log likelihood distribution, wherein the learning speech is applied to a probability model of the correct category to calculate a correct absolute log likelihood. Obtaining a correct answer absolute log likelihood distribution expressing what value spread the absolute log likelihood indicates, and the distribution of the likelihoods in each of the erroneous categories is a non-correct absolute log likelihood distribution. Calculating the non-correct absolute log likelihood by applying the above learning speech to the probability model of each incorrect category; and determining the value spread of each non-correct absolute log likelihood. Calculating a non-correct absolute log-likelihood distribution to be expressed.The likelihood obtained by applying the probability model of the correct category is a correct absolute log likelihood. 2. The voice quality evaluation according to claim 1, wherein the value indicating the probability and the value indicating the probability of being a non-correct answer are a correct absolute log likelihood probability density value and a non-correct absolute log likelihood probability density value, respectively. Method. 3. The step of obtaining the absolute score is a step of adding the difference between the correct absolute log likelihood probability density value and the non correct answer absolute log likelihood probability density value for each non correct answer category. 3. The voice quality evaluation method according to claim 2, wherein: 4. Using a plurality of linguistic categories of basic units such as phonemes, syllables, and words represented by a stochastic model related to acoustic features, and a likelihood of being a correct category calculated in advance from the stochastic model by a learning speech. And the distribution of the difference between the likelihoods of the wrong category (denoted as the relative distribution) and the distribution of the difference between the likelihoods of the wrong categories (denoted the non-correct relative distribution) are evaluated. Inputting a voice sample to be uttered and a basic unit sequence to be uttered corresponding to the voice, a step of decomposing the voice sample into basic units, and a probability model of a correct category for the decomposed voice basic unit. And calculating the likelihood of correct answer by applying the probability model of each erroneous category to the basic unit of the decomposed speech. Determining the difference between the correct likelihood and the non-correct likelihood as relative likelihood; and determining the relative likelihood from the relative likelihood and the correct relative distribution and the non-correct relative distribution as a correct category. Value indicating the likelihood that the difference between the wrong category and the wrong category (referred to as the relative correct likelihood value) and the value indicating the certainty that the relative likelihood is the difference between the wrong category and the wrong category (the non-correct answer) A voice quality evaluation method, comprising: a step of calculating a relative likelihood value); and a step of calculating a relative score to be evaluated from the two likelihood values. 5. The method according to claim 5, wherein said correct relative distribution is performed by applying said learning speech to a probability model of a correct category to calculate a correct absolute log likelihood; Calculating a correct absolute log likelihood; obtaining a difference between the correct absolute log likelihood and each of the non-correct absolute log likelihoods as a correct relative log likelihood; Obtaining a distribution representing whether the value spreads. The non-correct relative distribution is obtained by calculating a difference between the non-correct absolute log likelihoods as a non-correct relative log likelihood; and Obtaining a distribution expressing what value spread the log likelihood indicates. The correct likelihood is a correct absolute log likelihood, and the non-correct likelihood is a non-correct answer. Log likelihood, the relative likelihood is the difference between the correct absolute log likelihood and the non-correct absolute log likelihood, and the correct relative likelihood value and the non-correct relative likelihood value are respectively correct. The voice quality evaluation method according to claim 4, wherein the relative log likelihood probability density value and the non-correct relative log likelihood probability density value are used. 6. The step of calculating the relative score includes determining a difference between the correct relative log likelihood probability density value and the corresponding non-correct relative log likelihood probability density value by calculating a combination of error categories. 6. The voice quality evaluation method according to claim 5, further comprising the step of adding the difference between the two. 7. An overall score is obtained by linearly combining the absolute score obtained in any one of claims 1 to 3 and the corresponding relative score obtained in claim 4 to 7 as an evaluation result. A voice quality evaluation method characterized by the following. 8. Each of the log likelihood distributions is obtained by a step of obtaining a cumulative distribution function for the corresponding log likelihood value, and a step of calculating the probability distribution function from the sigmoid function. The voice quality evaluation method according to any one of claims 2, 3, 5, and 6. 9. The log likelihood distribution according to claim 2, wherein the log likelihood is calculated by calculating an average variance and applying a normal distribution to the corresponding log likelihood. Voice quality evaluation method described in Crab. 10. A recording medium on which a program for evaluating voice quality is recorded, wherein the program inputs a voice sample to be evaluated and a voice basic unit sequence to be uttered corresponding to the voice, Decomposing a speech sample into basic speech units; applying a correct category probability model to the decomposed basic speech units to calculate a correct absolute log likelihood; Calculating a correct absolute log likelihood probability density value from the absolute log likelihood distribution; and calculating a non-correct absolute log likelihood probability density value from the correct absolute log likelihood and a non-correct absolute log likelihood distribution obtained in advance. Calculating an absolute score representing the speech quality evaluation from the step and the correct absolute log likelihood probability density value and the non-correct absolute log likelihood probability density value A recording medium readable by a computer, comprising: 11. A recording medium on which a program for evaluating voice quality is recorded, wherein the program inputs a voice sample to be evaluated and a voice basic unit sequence to be uttered corresponding to the voice, Decomposing a voice sample into basic voice units; applying a correct category probability model to the decomposed basic voice unit to calculate a correct absolute log likelihood; Calculating the non-correct absolute log likelihood by applying the probability model of; calculating the difference between the correct absolute log likelihood and the non-correct absolute log likelihood to calculate the correct relative log likelihood; Calculating a correct relative log likelihood probability density value from a correct relative log likelihood distribution in which the correct relative log likelihood is obtained in advance; Calculating a non-correct relative log likelihood probability density value from the non-correct relative log likelihood distribution obtained in advance and the non-correct relative log likelihood probability density value and the non-correct relative log likelihood probability density value. Calculating a relative score representing the quality rating; and a computer readable recording medium. 12. A step of calculating a correct absolute log likelihood by applying a probability model of a correct category to the decomposed basic speech unit, and calculating a correct absolute log likelihood from the correct absolute log likelihood distribution and a previously calculated correct absolute log likelihood distribution. Calculating the absolute log likelihood probability density value; calculating the non-correct absolute log likelihood probability density value from the correct absolute log likelihood and the non-correct absolute log likelihood distribution obtained in advance; Calculating an absolute score representing quality evaluation from the likelihood probability density value and the non-correct absolute log likelihood probability density value; and obtaining a linear combination of the absolute score and the relative score to obtain a quality score. The recording medium according to claim 11, wherein the program is included. 13. A recording medium on which data used for evaluating voice quality is recorded, wherein the data is calculated by applying a voice basic unit of a learning voice to a probability model, and a correct absolute value representing a distribution of likelihood as a correct category. Log-likelihood distribution, non-correct absolute log-likelihood distribution indicating the distribution of likelihood as an incorrect category, and correct relative log indicating the distribution of the difference between likelihood as a correct category and likelihood as an incorrect category A computer-readable recording medium characterized by recording a likelihood distribution, and a non-correct log likelihood distribution representing a distribution of a difference between likelihoods each of which is an erroneous category.